Repository navigation
Per-section sync 3/5: per-section alignment engine (chroma + DTW per section, with confidence) #24
Description
Part of #8. Can start in parallel with the design decision; it only produces numbers, it does not change timestamps.
Reuse the chroma + DTW alignment from the current uniform sync redesign, but run it per section instead of once for the whole song.
Scope
- Use the arrangement's existing
sections[]markers (Intro / Verse / Chorus / Solo / ...) as region boundaries. - For each section, anchor on its opening notes against the audio for that region and produce a per-section
(factor, offset)or a set of anchor correspondences. - Produce a per-section confidence so a bad match on one section can be reported and skipped without affecting the others.
- Pure analysis: return results, do not modify note timestamps (baking is sub-issue 4).
Acceptance criteria
- On a synthetic or fixture song whose tempo changes at a section boundary, the per-section factors differ and each is closer to the truth than a single global factor (state the metric).
- A deliberately mismatched section gets a low confidence and does not change the others' results.
- Single-section songs reduce to the uniform model's result.
Out of scope
Applying the warp, the data model (sub-issue 2), and any UI (sub-issue 5).
Activity
pullfrog commentedon Oct 2, 2026
test: point the pinned uniform-sync baseline at the issue that retire…
test: pin the uniform sync's do-no-harm guarantee on a meter change (#22
pullfrog commentedon Oct 11, 2026
Per-section alignment engine (#24) is complete, tested, and open as PR #69 (vs
main, not merged).sourceTimes[0]), not at the boundary, so a section whose notes start inside/past its marker keeps the marker's place.audioStartsnapped to a real onset). The tight window is[audioStart, audioStart + slope·span + tailPad]; the lower edge is the snapped onset itself (previous-section content never enters), and the snap prefers the onset at-or-after the chained place (a stray prior-section click that edges a few tens of ms closer no longer hijacks the boundary).low-confidencewhile the sections before and after it stay byte-for-byte identical — and a separate regression pins the marker-place rule.Tests:
node tests/section_sync.test.mjs= 14/14;npm test= 401 pass / 0 fail;npm run lint= 0 errors (3 pre-existing warnings).Task list (2/6 completed)
big-pickle(free) ｜ 𝕏feat(sync): per-section alignment engine (#24)
pullfrog commentedon Oct 12, 2026
No new work needed — this issue is already implemented and merged.
main; HEAD (4e1acb9,feat(sync): per-section alignment engine (#24)) containssrc/section-sync.jsandtests/section_sync.test.mjs.node tests/section_sync.test.mjs= 14/14 andnpm test= 401 pass / 0 fail. All three acceptance criteria (per-section factors beat the best single global fit; a mismatched section reads low confidence and leaves its neighbours byte-for-byte identical; a single section reduces to the uniform factor) are covered.Closing the issue as completed, since the tracked work is on
main.big-pickle(free) ｜ 𝕏feat(sync): per-section sync results UI (#26) (#71)